Research map / Label-efficient and robust learning / Self-supervised and contrastive
Sketch-BERT: Learning Sketch Bidirectional Encoder Representation From Transformers by Self-Supervised Learning of Sketch Gestalt
Hangyu Lin, Yanwei Fu, Xiangyang Xue, Yu-Gang Jiang
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sketch Gestalt self-supervised pre-training (ablation) | Sketch-BERT · QuickDraw classification | Top-1 accuracy: 83.10 -> 88.30 | Improved |
| add sketch Gestalt self-supervised pre-training (ablation) | Sketch-BERT · TU-Berlin classification | Top-1 accuracy: 54.20 -> 76.30 | Improved |
| add sketch Gestalt self-supervised pre-training (ablation) | Sketch-BERT · QuickDraw retrieval | mAP: 55.10 -> 78.87 | Improved |
| add sketch Gestalt self-supervised pre-training (ablation) | Sketch-BERT · TU-Berlin retrieval | mAP: 24.14 -> 41.54 | Improved |
| use full position-and-state sketch Gestalt masking rather than BERT-style random masking (ablation) | Sketch-BERT pre-training · QuickDraw classification | Top-1 accuracy: 86.51 -> 88.30 | Improved |
| mask both positions and states rather than positions only (ablation) | Sketch-BERT pre-training · QuickDraw classification | Top-1 accuracy: 87.37 -> 88.30 | Improved |
| mask positions and states rather than states only (ablation) | Sketch-BERT pre-training · QuickDraw retrieval | Top-1 accuracy: 81.87 -> 85.47 | Improved |
| increase pre-training samples per class from 5K to 70K (ablation) | Sketch-BERT · QuickDraw classification | Top-1 accuracy: 85.73 -> 88.30 | Improved |
| increase pre-training categories from 100 to 200 at 5K samples per class (ablation) | Sketch-BERT · QuickDraw classification | Top-1 accuracy: 85.82 -> 84.89 | Hurt |
| increase transformer depth from 6 to 12 layers at hidden size 256 (ablation) | Sketch-BERT · QuickDraw classification | Top-1 accuracy: 84.83 -> 86.34 | Improved |
| use 12-16-1024 instead of 8-12-768 architecture (ablation) | Sketch-BERT · QuickDraw classification | Top-1 accuracy: 88.30 -> 85.31 | Hurt |
| use Sketch-BERT rather than TCNet | sketch retrieval · QuickDraw | mAP: 76.38 -> 78.87 | Improved |
Closest papers by meaning
- Sketchformer: Transformer-Based Representation for Sketched Structure (CVPR 2020 · similarity 0.91)
- SketchINR: A First Look into Sketches as Implicit Neural Representations (CVPR 2024 · similarity 0.89)
- SketchEmbedNet: Learning Novel Concepts by Imitating Drawings (ICML 2021 · similarity 0.88)
- Enhance Sketch Recognition’s Explainability via Semantic Component-Level Parsing (AAAI 2024 · similarity 0.86)
- Open Vocabulary Semantic Scene Sketch Understanding (CVPR 2024 · similarity 0.86)
- SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
